module/zfs/dsl_scrub.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
  23  * Use is subject to license terms.
  24  */
  25
  26 #include <sys/dsl_pool.h>
  27 #include <sys/dsl_dataset.h>
  28 #include <sys/dsl_prop.h>
  29 #include <sys/dsl_dir.h>
  30 #include <sys/dsl_synctask.h>
  31 #include <sys/dnode.h>
  32 #include <sys/dmu_tx.h>
  33 #include <sys/dmu_objset.h>
  34 #include <sys/arc.h>
  35 #include <sys/zap.h>
  36 #include <sys/zio.h>
  37 #include <sys/zfs_context.h>
  38 #include <sys/fs/zfs.h>
  39 #include <sys/zfs_znode.h>
  40 #include <sys/spa_impl.h>
  41 #include <sys/vdev_impl.h>
  42 #include <sys/zil_impl.h>
  43
  44 typedef int (scrub_cb_t)(dsl_pool_t *, const blkptr_t *, const zbookmark_t *);
  45
  46 static scrub_cb_t dsl_pool_scrub_clean_cb;
  47 static dsl_syncfunc_t dsl_pool_scrub_cancel_sync;
  48
  49 int zfs_scrub_min_time = 1; /* scrub for at least 1 sec each txg */
  50 int zfs_resilver_min_time = 3; /* resilver for at least 3 sec each txg */
  51 boolean_t zfs_no_scrub_io = B_FALSE; /* set to disable scrub i/o */
  52
  53 extern int zfs_txg_timeout;
  54
  55 static scrub_cb_t *scrub_funcs[SCRUB_FUNC_NUMFUNCS] = {
  56         NULL,
  57         dsl_pool_scrub_clean_cb
  58 };
  59
  60 #define SET_BOOKMARK(zb, objset, object, level, blkid)  \
  61 {                                                       \
  62         (zb)->zb_objset = objset;                       \
  63         (zb)->zb_object = object;                       \
  64         (zb)->zb_level = level;                         \
  65         (zb)->zb_blkid = blkid;                         \
  66 }
  67
  68 /* ARGSUSED */
  69 static void
  70 dsl_pool_scrub_setup_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
  71 {
  72         dsl_pool_t *dp = arg1;
  73         enum scrub_func *funcp = arg2;
  74         dmu_object_type_t ot = 0;
  75         boolean_t complete = B_FALSE;
  76
  77         dsl_pool_scrub_cancel_sync(dp, &complete, cr, tx);
  78
  79         ASSERT(dp->dp_scrub_func == SCRUB_FUNC_NONE);
  80         ASSERT(*funcp > SCRUB_FUNC_NONE);
  81         ASSERT(*funcp < SCRUB_FUNC_NUMFUNCS);
  82
  83         dp->dp_scrub_min_txg = 0;
  84         dp->dp_scrub_max_txg = tx->tx_txg;
  85
  86         if (*funcp == SCRUB_FUNC_CLEAN) {
  87                 vdev_t *rvd = dp->dp_spa->spa_root_vdev;
  88
  89                 /* rewrite all disk labels */
  90                 vdev_config_dirty(rvd);
  91
  92                 if (vdev_resilver_needed(rvd,
  93                     &dp->dp_scrub_min_txg, &dp->dp_scrub_max_txg)) {
  94                         spa_event_notify(dp->dp_spa, NULL,
  95                             ESC_ZFS_RESILVER_START);
  96                         dp->dp_scrub_max_txg = MIN(dp->dp_scrub_max_txg,
  97                             tx->tx_txg);
  98                 }
  99
 100                 /* zero out the scrub stats in all vdev_stat_t's */
 101                 vdev_scrub_stat_update(rvd,
 102                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 103                     POOL_SCRUB_EVERYTHING, B_FALSE);
 104
 105                 dp->dp_spa->spa_scrub_started = B_TRUE;
 106         }
 107
 108         /* back to the generic stuff */
 109
 110         if (dp->dp_blkstats == NULL) {
 111                 dp->dp_blkstats =
 112                     kmem_alloc(sizeof (zfs_all_blkstats_t), KM_SLEEP);
 113         }
 114         bzero(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
 115
 116         if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB)
 117                 ot = DMU_OT_ZAP_OTHER;
 118
 119         dp->dp_scrub_func = *funcp;
 120         dp->dp_scrub_queue_obj = zap_create(dp->dp_meta_objset,
 121             ot ? ot : DMU_OT_SCRUB_QUEUE, DMU_OT_NONE, 0, tx);
 122         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 123         dp->dp_scrub_restart = B_FALSE;
 124         dp->dp_spa->spa_scrub_errors = 0;
 125
 126         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 127             DMU_POOL_SCRUB_FUNC, sizeof (uint32_t), 1,
 128             &dp->dp_scrub_func, tx));
 129         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 130             DMU_POOL_SCRUB_QUEUE, sizeof (uint64_t), 1,
 131             &dp->dp_scrub_queue_obj, tx));
 132         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 133             DMU_POOL_SCRUB_MIN_TXG, sizeof (uint64_t), 1,
 134             &dp->dp_scrub_min_txg, tx));
 135         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 136             DMU_POOL_SCRUB_MAX_TXG, sizeof (uint64_t), 1,
 137             &dp->dp_scrub_max_txg, tx));
 138         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 139             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 140             &dp->dp_scrub_bookmark, tx));
 141         VERIFY(0 == zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 142             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 143             &dp->dp_spa->spa_scrub_errors, tx));
 144
 145         spa_history_internal_log(LOG_POOL_SCRUB, dp->dp_spa, tx, cr,
 146             "func=%u mintxg=%llu maxtxg=%llu",
 147             *funcp, dp->dp_scrub_min_txg, dp->dp_scrub_max_txg);
 148 }
 149
 150 int
 151 dsl_pool_scrub_setup(dsl_pool_t *dp, enum scrub_func func)
 152 {
 153         return (dsl_sync_task_do(dp, NULL,
 154             dsl_pool_scrub_setup_sync, dp, &func, 0));
 155 }
 156
 157 /* ARGSUSED */
 158 static void
 159 dsl_pool_scrub_cancel_sync(void *arg1, void *arg2, cred_t *cr, dmu_tx_t *tx)
 160 {
 161         dsl_pool_t *dp = arg1;
 162         boolean_t *completep = arg2;
 163
 164         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 165                 return;
 166
 167         mutex_enter(&dp->dp_scrub_cancel_lock);
 168
 169         if (dp->dp_scrub_restart) {
 170                 dp->dp_scrub_restart = B_FALSE;
 171                 *completep = B_FALSE;
 172         }
 173
 174         /* XXX this is scrub-clean specific */
 175         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 176         while (dp->dp_spa->spa_scrub_inflight > 0) {
 177                 cv_wait(&dp->dp_spa->spa_scrub_io_cv,
 178                     &dp->dp_spa->spa_scrub_lock);
 179         }
 180         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 181         dp->dp_spa->spa_scrub_started = B_FALSE;
 182         dp->dp_spa->spa_scrub_active = B_FALSE;
 183
 184         dp->dp_scrub_func = SCRUB_FUNC_NONE;
 185         VERIFY(0 == dmu_object_free(dp->dp_meta_objset,
 186             dp->dp_scrub_queue_obj, tx));
 187         dp->dp_scrub_queue_obj = 0;
 188         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 189
 190         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 191             DMU_POOL_SCRUB_QUEUE, tx));
 192         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 193             DMU_POOL_SCRUB_MIN_TXG, tx));
 194         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 195             DMU_POOL_SCRUB_MAX_TXG, tx));
 196         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 197             DMU_POOL_SCRUB_BOOKMARK, tx));
 198         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 199             DMU_POOL_SCRUB_FUNC, tx));
 200         VERIFY(0 == zap_remove(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
 201             DMU_POOL_SCRUB_ERRORS, tx));
 202
 203         spa_history_internal_log(LOG_POOL_SCRUB_DONE, dp->dp_spa, tx, cr,
 204             "complete=%u", *completep);
 205
 206         /* below is scrub-clean specific */
 207         vdev_scrub_stat_update(dp->dp_spa->spa_root_vdev, POOL_SCRUB_NONE,
 208             *completep);
 209         /*
 210          * If the scrub/resilver completed, update all DTLs to reflect this.
 211          * Whether it succeeded or not, vacate all temporary scrub DTLs.
 212          */
 213         vdev_dtl_reassess(dp->dp_spa->spa_root_vdev, tx->tx_txg,
 214             *completep ? dp->dp_scrub_max_txg : 0, B_TRUE);
 215         if (dp->dp_scrub_min_txg && *completep)
 216                 spa_event_notify(dp->dp_spa, NULL, ESC_ZFS_RESILVER_FINISH);
 217         spa_errlog_rotate(dp->dp_spa);
 218
 219         /*
 220          * We may have finished replacing a device.
 221          * Let the async thread assess this and handle the detach.
 222          */
 223         spa_async_request(dp->dp_spa, SPA_ASYNC_RESILVER_DONE);
 224
 225         dp->dp_scrub_min_txg = dp->dp_scrub_max_txg = 0;
 226         mutex_exit(&dp->dp_scrub_cancel_lock);
 227 }
 228
 229 int
 230 dsl_pool_scrub_cancel(dsl_pool_t *dp)
 231 {
 232         boolean_t complete = B_FALSE;
 233
 234         return (dsl_sync_task_do(dp, NULL,
 235             dsl_pool_scrub_cancel_sync, dp, &complete, 3));
 236 }
 237
 238 int
 239 dsl_free(zio_t *pio, dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp,
 240     zio_done_func_t *done, void *private, uint32_t arc_flags)
 241 {
 242         /*
 243          * This function will be used by bp-rewrite wad to intercept frees.
 244          */
 245         return (arc_free(pio, dp->dp_spa, txg, (blkptr_t *)bpp,
 246             done, private, arc_flags));
 247 }
 248
 249 static boolean_t
 250 bookmark_is_zero(const zbookmark_t *zb)
 251 {
 252         return (zb->zb_objset == 0 && zb->zb_object == 0 &&
 253             zb->zb_level == 0 && zb->zb_blkid == 0);
 254 }
 255
 256 /* dnp is the dnode for zb1->zb_object */
 257 static boolean_t
 258 bookmark_is_before(dnode_phys_t *dnp, const zbookmark_t *zb1,
 259     const zbookmark_t *zb2)
 260 {
 261         uint64_t zb1nextL0, zb2thisobj;
 262
 263         ASSERT(zb1->zb_objset == zb2->zb_objset);
 264         ASSERT(zb1->zb_object != -1ULL);
 265         ASSERT(zb2->zb_level == 0);
 266
 267         /*
 268          * A bookmark in the deadlist is considered to be after
 269          * everything else.
 270          */
 271         if (zb2->zb_object == -1ULL)
 272                 return (B_TRUE);
 273
 274         /* The objset_phys_t isn't before anything. */
 275         if (dnp == NULL)
 276                 return (B_FALSE);
 277
 278         zb1nextL0 = (zb1->zb_blkid + 1) <<
 279             ((zb1->zb_level) * (dnp->dn_indblkshift - SPA_BLKPTRSHIFT));
 280
 281         zb2thisobj = zb2->zb_object ? zb2->zb_object :
 282             zb2->zb_blkid << (DNODE_BLOCK_SHIFT - DNODE_SHIFT);
 283
 284         if (zb1->zb_object == 0) {
 285                 uint64_t nextobj = zb1nextL0 *
 286                     (dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT) >> DNODE_SHIFT;
 287                 return (nextobj <= zb2thisobj);
 288         }
 289
 290         if (zb1->zb_object < zb2thisobj)
 291                 return (B_TRUE);
 292         if (zb1->zb_object > zb2thisobj)
 293                 return (B_FALSE);
 294         if (zb2->zb_object == 0)
 295                 return (B_FALSE);
 296         return (zb1nextL0 <= zb2->zb_blkid);
 297 }
 298
 299 static boolean_t
 300 scrub_pause(dsl_pool_t *dp, const zbookmark_t *zb)
 301 {
 302         int elapsed_ticks;
 303         int mintime;
 304
 305         if (dp->dp_scrub_pausing)
 306                 return (B_TRUE); /* we're already pausing */
 307
 308         if (!bookmark_is_zero(&dp->dp_scrub_bookmark))
 309                 return (B_FALSE); /* we're resuming */
 310
 311         /* We only know how to resume from level-0 blocks. */
 312         if (zb->zb_level != 0)
 313                 return (B_FALSE);
 314
 315         mintime = dp->dp_scrub_isresilver ? zfs_resilver_min_time :
 316             zfs_scrub_min_time;
 317         elapsed_ticks = lbolt64 - dp->dp_scrub_start_time;
 318         if (elapsed_ticks > hz * zfs_txg_timeout ||
 319             (elapsed_ticks > hz * mintime && txg_sync_waiting(dp))) {
 320                 dprintf("pausing at %llx/%llx/%llx/%llx\n",
 321                     (longlong_t)zb->zb_objset, (longlong_t)zb->zb_object,
 322                     (longlong_t)zb->zb_level, (longlong_t)zb->zb_blkid);
 323                 dp->dp_scrub_pausing = B_TRUE;
 324                 dp->dp_scrub_bookmark = *zb;
 325                 return (B_TRUE);
 326         }
 327         return (B_FALSE);
 328 }
 329
 330 typedef struct zil_traverse_arg {
 331         dsl_pool_t      *zta_dp;
 332         zil_header_t    *zta_zh;
 333 } zil_traverse_arg_t;
 334
 335 /* ARGSUSED */
 336 static void
 337 traverse_zil_block(zilog_t *zilog, blkptr_t *bp, void *arg, uint64_t claim_txg)
 338 {
 339         zil_traverse_arg_t *zta = arg;
 340         dsl_pool_t *dp = zta->zta_dp;
 341         zil_header_t *zh = zta->zta_zh;
 342         zbookmark_t zb;
 343
 344         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 345                 return;
 346
 347         if (claim_txg == 0 && bp->blk_birth >= spa_first_txg(dp->dp_spa))
 348                 return;
 349
 350         zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 351         zb.zb_object = 0;
 352         zb.zb_level = -1;
 353         zb.zb_blkid = bp->blk_cksum.zc_word[ZIL_ZC_SEQ];
 354         VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 355 }
 356
 357 /* ARGSUSED */
 358 static void
 359 traverse_zil_record(zilog_t *zilog, lr_t *lrc, void *arg, uint64_t claim_txg)
 360 {
 361         if (lrc->lrc_txtype == TX_WRITE) {
 362                 zil_traverse_arg_t *zta = arg;
 363                 dsl_pool_t *dp = zta->zta_dp;
 364                 zil_header_t *zh = zta->zta_zh;
 365                 lr_write_t *lr = (lr_write_t *)lrc;
 366                 blkptr_t *bp = &lr->lr_blkptr;
 367                 zbookmark_t zb;
 368
 369                 if (bp->blk_birth <= dp->dp_scrub_min_txg)
 370                         return;
 371
 372                 if (claim_txg == 0 || bp->blk_birth < claim_txg)
 373                         return;
 374
 375                 zb.zb_objset = zh->zh_log.blk_cksum.zc_word[ZIL_ZC_OBJSET];
 376                 zb.zb_object = lr->lr_foid;
 377                 zb.zb_level = BP_GET_LEVEL(bp);
 378                 zb.zb_blkid = lr->lr_offset / BP_GET_LSIZE(bp);
 379                 VERIFY(0 == scrub_funcs[dp->dp_scrub_func](dp, bp, &zb));
 380         }
 381 }
 382
 383 static void
 384 traverse_zil(dsl_pool_t *dp, zil_header_t *zh)
 385 {
 386         uint64_t claim_txg = zh->zh_claim_txg;
 387         zil_traverse_arg_t zta = { dp, zh };
 388         zilog_t *zilog;
 389
 390         /*
 391          * We only want to visit blocks that have been claimed but not yet
 392          * replayed (or, in read-only mode, blocks that *would* be claimed).
 393          */
 394         if (claim_txg == 0 && spa_writeable(dp->dp_spa))
 395                 return;
 396
 397         zilog = zil_alloc(dp->dp_meta_objset, zh);
 398
 399         (void) zil_parse(zilog, traverse_zil_block, traverse_zil_record, &zta,
 400             claim_txg);
 401
 402         zil_free(zilog);
 403 }
 404
 405 static void
 406 scrub_visitbp(dsl_pool_t *dp, dnode_phys_t *dnp,
 407     arc_buf_t *pbuf, blkptr_t *bp, const zbookmark_t *zb)
 408 {
 409         int err;
 410         arc_buf_t *buf = NULL;
 411
 412         if (bp->blk_birth <= dp->dp_scrub_min_txg)
 413                 return;
 414
 415         if (scrub_pause(dp, zb))
 416                 return;
 417
 418         if (!bookmark_is_zero(&dp->dp_scrub_bookmark)) {
 419                 /*
 420                  * If we already visited this bp & everything below (in
 421                  * a prior txg), don't bother doing it again.
 422                  */
 423                 if (bookmark_is_before(dnp, zb, &dp->dp_scrub_bookmark))
 424                         return;
 425
 426                 /*
 427                  * If we found the block we're trying to resume from, or
 428                  * we went past it to a different object, zero it out to
 429                  * indicate that it's OK to start checking for pausing
 430                  * again.
 431                  */
 432                 if (bcmp(zb, &dp->dp_scrub_bookmark, sizeof (*zb)) == 0 ||
 433                     zb->zb_object > dp->dp_scrub_bookmark.zb_object) {
 434                         dprintf("resuming at %llx/%llx/%llx/%llx\n",
 435                             (longlong_t)zb->zb_objset,
 436                             (longlong_t)zb->zb_object,
 437                             (longlong_t)zb->zb_level,
 438                             (longlong_t)zb->zb_blkid);
 439                         bzero(&dp->dp_scrub_bookmark, sizeof (*zb));
 440                 }
 441         }
 442
 443         if (BP_GET_LEVEL(bp) > 0) {
 444                 uint32_t flags = ARC_WAIT;
 445                 int i;
 446                 blkptr_t *cbp;
 447                 int epb = BP_GET_LSIZE(bp) >> SPA_BLKPTRSHIFT;
 448
 449                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 450                     arc_getbuf_func, &buf,
 451                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 452                 if (err) {
 453                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 454                         dp->dp_spa->spa_scrub_errors++;
 455                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 456                         return;
 457                 }
 458                 cbp = buf->b_data;
 459
 460                 for (i = 0; i < epb; i++, cbp++) {
 461                         zbookmark_t czb;
 462
 463                         SET_BOOKMARK(&czb, zb->zb_objset, zb->zb_object,
 464                             zb->zb_level - 1,
 465                             zb->zb_blkid * epb + i);
 466                         scrub_visitbp(dp, dnp, buf, cbp, &czb);
 467                 }
 468         } else if (BP_GET_TYPE(bp) == DMU_OT_DNODE) {
 469                 uint32_t flags = ARC_WAIT;
 470                 dnode_phys_t *child_dnp;
 471                 int i, j;
 472                 int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
 473
 474                 err = arc_read(NULL, dp->dp_spa, bp, pbuf,
 475                     arc_getbuf_func, &buf,
 476                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 477                 if (err) {
 478                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 479                         dp->dp_spa->spa_scrub_errors++;
 480                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 481                         return;
 482                 }
 483                 child_dnp = buf->b_data;
 484
 485                 for (i = 0; i < epb; i++, child_dnp++) {
 486                         for (j = 0; j < child_dnp->dn_nblkptr; j++) {
 487                                 zbookmark_t czb;
 488
 489                                 SET_BOOKMARK(&czb, zb->zb_objset,
 490                                     zb->zb_blkid * epb + i,
 491                                     child_dnp->dn_nlevels - 1, j);
 492                                 scrub_visitbp(dp, child_dnp, buf,
 493                                     &child_dnp->dn_blkptr[j], &czb);
 494                         }
 495                 }
 496         } else if (BP_GET_TYPE(bp) == DMU_OT_OBJSET) {
 497                 uint32_t flags = ARC_WAIT;
 498                 objset_phys_t *osp;
 499                 int j;
 500
 501                 err = arc_read_nolock(NULL, dp->dp_spa, bp,
 502                     arc_getbuf_func, &buf,
 503                     ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL, &flags, zb);
 504                 if (err) {
 505                         mutex_enter(&dp->dp_spa->spa_scrub_lock);
 506                         dp->dp_spa->spa_scrub_errors++;
 507                         mutex_exit(&dp->dp_spa->spa_scrub_lock);
 508                         return;
 509                 }
 510
 511                 osp = buf->b_data;
 512
 513                 traverse_zil(dp, &osp->os_zil_header);
 514
 515                 for (j = 0; j < osp->os_meta_dnode.dn_nblkptr; j++) {
 516                         zbookmark_t czb;
 517
 518                         SET_BOOKMARK(&czb, zb->zb_objset, 0,
 519                             osp->os_meta_dnode.dn_nlevels - 1, j);
 520                         scrub_visitbp(dp, &osp->os_meta_dnode, buf,
 521                             &osp->os_meta_dnode.dn_blkptr[j], &czb);
 522                 }
 523         }
 524
 525         (void) scrub_funcs[dp->dp_scrub_func](dp, bp, zb);
 526         if (buf)
 527                 (void) arc_buf_remove_ref(buf, &buf);
 528 }
 529
 530 static void
 531 scrub_visit_rootbp(dsl_pool_t *dp, dsl_dataset_t *ds, blkptr_t *bp)
 532 {
 533         zbookmark_t zb;
 534
 535         SET_BOOKMARK(&zb, ds ? ds->ds_object : 0, 0, -1, 0);
 536         scrub_visitbp(dp, NULL, NULL, bp, &zb);
 537 }
 538
 539 void
 540 dsl_pool_ds_destroyed(dsl_dataset_t *ds, dmu_tx_t *tx)
 541 {
 542         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 543
 544         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 545                 return;
 546
 547         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 548                 SET_BOOKMARK(&dp->dp_scrub_bookmark, -1, 0, 0, 0);
 549         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 550             ds->ds_object, tx) != 0) {
 551                 return;
 552         }
 553
 554         if (ds->ds_phys->ds_next_snap_obj != 0) {
 555                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 556                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 557         }
 558         ASSERT3U(ds->ds_phys->ds_num_children, <=, 1);
 559 }
 560
 561 void
 562 dsl_pool_ds_snapshotted(dsl_dataset_t *ds, dmu_tx_t *tx)
 563 {
 564         dsl_pool_t *dp = ds->ds_dir->dd_pool;
 565
 566         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 567                 return;
 568
 569         ASSERT(ds->ds_phys->ds_prev_snap_obj != 0);
 570
 571         if (dp->dp_scrub_bookmark.zb_objset == ds->ds_object) {
 572                 dp->dp_scrub_bookmark.zb_objset =
 573                     ds->ds_phys->ds_prev_snap_obj;
 574         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 575             ds->ds_object, tx) == 0) {
 576                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 577                     ds->ds_phys->ds_prev_snap_obj, tx) == 0);
 578         }
 579 }
 580
 581 void
 582 dsl_pool_ds_clone_swapped(dsl_dataset_t *ds1, dsl_dataset_t *ds2, dmu_tx_t *tx)
 583 {
 584         dsl_pool_t *dp = ds1->ds_dir->dd_pool;
 585
 586         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 587                 return;
 588
 589         if (dp->dp_scrub_bookmark.zb_objset == ds1->ds_object) {
 590                 dp->dp_scrub_bookmark.zb_objset = ds2->ds_object;
 591         } else if (dp->dp_scrub_bookmark.zb_objset == ds2->ds_object) {
 592                 dp->dp_scrub_bookmark.zb_objset = ds1->ds_object;
 593         }
 594
 595         if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 596             ds1->ds_object, tx) == 0) {
 597                 int err = zap_add_int(dp->dp_meta_objset,
 598                     dp->dp_scrub_queue_obj, ds2->ds_object, tx);
 599                 VERIFY(err == 0 || err == EEXIST);
 600                 if (err == EEXIST) {
 601                         /* Both were there to begin with */
 602                         VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 603                             dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 604                 }
 605         } else if (zap_remove_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 606             ds2->ds_object, tx) == 0) {
 607                 VERIFY(0 == zap_add_int(dp->dp_meta_objset,
 608                     dp->dp_scrub_queue_obj, ds1->ds_object, tx));
 609         }
 610 }
 611
 612 struct enqueue_clones_arg {
 613         dmu_tx_t *tx;
 614         uint64_t originobj;
 615 };
 616
 617 /* ARGSUSED */
 618 static int
 619 enqueue_clones_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 620 {
 621         struct enqueue_clones_arg *eca = arg;
 622         dsl_dataset_t *ds;
 623         int err;
 624         dsl_pool_t *dp;
 625
 626         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 627         if (err)
 628                 return (err);
 629         dp = ds->ds_dir->dd_pool;
 630
 631         if (ds->ds_dir->dd_phys->dd_origin_obj == eca->originobj) {
 632                 while (ds->ds_phys->ds_prev_snap_obj != eca->originobj) {
 633                         dsl_dataset_t *prev;
 634                         err = dsl_dataset_hold_obj(dp,
 635                             ds->ds_phys->ds_prev_snap_obj, FTAG, &prev);
 636
 637                         dsl_dataset_rele(ds, FTAG);
 638                         if (err)
 639                                 return (err);
 640                         ds = prev;
 641                 }
 642                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 643                     ds->ds_object, eca->tx) == 0);
 644         }
 645         dsl_dataset_rele(ds, FTAG);
 646         return (0);
 647 }
 648
 649 static void
 650 scrub_visitds(dsl_pool_t *dp, uint64_t dsobj, dmu_tx_t *tx)
 651 {
 652         dsl_dataset_t *ds;
 653         uint64_t min_txg_save;
 654
 655         VERIFY3U(0, ==, dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
 656
 657         /*
 658          * Iterate over the bps in this ds.
 659          */
 660         min_txg_save = dp->dp_scrub_min_txg;
 661         dp->dp_scrub_min_txg =
 662             MAX(dp->dp_scrub_min_txg, ds->ds_phys->ds_prev_snap_txg);
 663         scrub_visit_rootbp(dp, ds, &ds->ds_phys->ds_bp);
 664         dp->dp_scrub_min_txg = min_txg_save;
 665
 666         if (dp->dp_scrub_pausing)
 667                 goto out;
 668
 669         /*
 670          * Add descendent datasets to work queue.
 671          */
 672         if (ds->ds_phys->ds_next_snap_obj != 0) {
 673                 VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 674                     ds->ds_phys->ds_next_snap_obj, tx) == 0);
 675         }
 676         if (ds->ds_phys->ds_num_children > 1) {
 677                 if (spa_version(dp->dp_spa) < SPA_VERSION_DSL_SCRUB) {
 678                         struct enqueue_clones_arg eca;
 679                         eca.tx = tx;
 680                         eca.originobj = ds->ds_object;
 681
 682                         (void) dmu_objset_find_spa(ds->ds_dir->dd_pool->dp_spa,
 683                             NULL, enqueue_clones_cb, &eca, DS_FIND_CHILDREN);
 684                 } else {
 685                         VERIFY(zap_join(dp->dp_meta_objset,
 686                             ds->ds_phys->ds_next_clones_obj,
 687                             dp->dp_scrub_queue_obj, tx) == 0);
 688                 }
 689         }
 690
 691 out:
 692         dsl_dataset_rele(ds, FTAG);
 693 }
 694
 695 /* ARGSUSED */
 696 static int
 697 enqueue_cb(spa_t *spa, uint64_t dsobj, const char *dsname, void *arg)
 698 {
 699         dmu_tx_t *tx = arg;
 700         dsl_dataset_t *ds;
 701         int err;
 702         dsl_pool_t *dp;
 703
 704         err = dsl_dataset_hold_obj(spa->spa_dsl_pool, dsobj, FTAG, &ds);
 705         if (err)
 706                 return (err);
 707
 708         dp = ds->ds_dir->dd_pool;
 709
 710         while (ds->ds_phys->ds_prev_snap_obj != 0) {
 711                 dsl_dataset_t *prev;
 712                 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
 713                     FTAG, &prev);
 714                 if (err) {
 715                         dsl_dataset_rele(ds, FTAG);
 716                         return (err);
 717                 }
 718
 719                 /*
 720                  * If this is a clone, we don't need to worry about it for now.
 721                  */
 722                 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object) {
 723                         dsl_dataset_rele(ds, FTAG);
 724                         dsl_dataset_rele(prev, FTAG);
 725                         return (0);
 726                 }
 727                 dsl_dataset_rele(ds, FTAG);
 728                 ds = prev;
 729         }
 730
 731         VERIFY(zap_add_int(dp->dp_meta_objset, dp->dp_scrub_queue_obj,
 732             ds->ds_object, tx) == 0);
 733         dsl_dataset_rele(ds, FTAG);
 734         return (0);
 735 }
 736
 737 void
 738 dsl_pool_scrub_sync(dsl_pool_t *dp, dmu_tx_t *tx)
 739 {
 740         spa_t *spa = dp->dp_spa;
 741         zap_cursor_t zc;
 742         zap_attribute_t za;
 743         boolean_t complete = B_TRUE;
 744
 745         if (dp->dp_scrub_func == SCRUB_FUNC_NONE)
 746                 return;
 747
 748         /*
 749          * If the pool is not loaded, or is trying to unload, leave it alone.
 750          */
 751         if (spa->spa_load_state != SPA_LOAD_NONE || spa_shutting_down(spa))
 752                 return;
 753
 754         if (dp->dp_scrub_restart) {
 755                 enum scrub_func func = dp->dp_scrub_func;
 756                 dp->dp_scrub_restart = B_FALSE;
 757                 dsl_pool_scrub_setup_sync(dp, &func, kcred, tx);
 758         }
 759
 760         if (spa->spa_root_vdev->vdev_stat.vs_scrub_type == 0) {
 761                 /*
 762                  * We must have resumed after rebooting; reset the vdev
 763                  * stats to know that we're doing a scrub (although it
 764                  * will think we're just starting now).
 765                  */
 766                 vdev_scrub_stat_update(spa->spa_root_vdev,
 767                     dp->dp_scrub_min_txg ? POOL_SCRUB_RESILVER :
 768                     POOL_SCRUB_EVERYTHING, B_FALSE);
 769         }
 770
 771         dp->dp_scrub_pausing = B_FALSE;
 772         dp->dp_scrub_start_time = lbolt64;
 773         dp->dp_scrub_isresilver = (dp->dp_scrub_min_txg != 0);
 774         spa->spa_scrub_active = B_TRUE;
 775
 776         if (dp->dp_scrub_bookmark.zb_objset == 0) {
 777                 /* First do the MOS & ORIGIN */
 778                 scrub_visit_rootbp(dp, NULL, &dp->dp_meta_rootbp);
 779                 if (dp->dp_scrub_pausing)
 780                         goto out;
 781
 782                 if (spa_version(spa) < SPA_VERSION_DSL_SCRUB) {
 783                         VERIFY(0 == dmu_objset_find_spa(spa,
 784                             NULL, enqueue_cb, tx, DS_FIND_CHILDREN));
 785                 } else {
 786                         scrub_visitds(dp, dp->dp_origin_snap->ds_object, tx);
 787                 }
 788                 ASSERT(!dp->dp_scrub_pausing);
 789         } else if (dp->dp_scrub_bookmark.zb_objset != -1ULL) {
 790                 /*
 791                  * If we were paused, continue from here.  Note if the
 792                  * ds we were paused on was deleted, the zb_objset will
 793                  * be -1, so we will skip this and find a new objset
 794                  * below.
 795                  */
 796                 scrub_visitds(dp, dp->dp_scrub_bookmark.zb_objset, tx);
 797                 if (dp->dp_scrub_pausing)
 798                         goto out;
 799         }
 800
 801         /*
 802          * In case we were paused right at the end of the ds, zero the
 803          * bookmark so we don't think that we're still trying to resume.
 804          */
 805         bzero(&dp->dp_scrub_bookmark, sizeof (zbookmark_t));
 806
 807         /* keep pulling things out of the zap-object-as-queue */
 808         while (zap_cursor_init(&zc, dp->dp_meta_objset, dp->dp_scrub_queue_obj),
 809             zap_cursor_retrieve(&zc, &za) == 0) {
 810                 VERIFY(0 == zap_remove(dp->dp_meta_objset,
 811                     dp->dp_scrub_queue_obj, za.za_name, tx));
 812                 scrub_visitds(dp, za.za_first_integer, tx);
 813                 if (dp->dp_scrub_pausing)
 814                         break;
 815                 zap_cursor_fini(&zc);
 816         }
 817         zap_cursor_fini(&zc);
 818         if (dp->dp_scrub_pausing)
 819                 goto out;
 820
 821         /* done. */
 822
 823         dsl_pool_scrub_cancel_sync(dp, &complete, kcred, tx);
 824         return;
 825 out:
 826         VERIFY(0 == zap_update(dp->dp_meta_objset,
 827             DMU_POOL_DIRECTORY_OBJECT,
 828             DMU_POOL_SCRUB_BOOKMARK, sizeof (uint64_t), 4,
 829             &dp->dp_scrub_bookmark, tx));
 830         VERIFY(0 == zap_update(dp->dp_meta_objset,
 831             DMU_POOL_DIRECTORY_OBJECT,
 832             DMU_POOL_SCRUB_ERRORS, sizeof (uint64_t), 1,
 833             &spa->spa_scrub_errors, tx));
 834
 835         /* XXX this is scrub-clean specific */
 836         mutex_enter(&spa->spa_scrub_lock);
 837         while (spa->spa_scrub_inflight > 0)
 838                 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
 839         mutex_exit(&spa->spa_scrub_lock);
 840 }
 841
 842 void
 843 dsl_pool_scrub_restart(dsl_pool_t *dp)
 844 {
 845         mutex_enter(&dp->dp_scrub_cancel_lock);
 846         dp->dp_scrub_restart = B_TRUE;
 847         mutex_exit(&dp->dp_scrub_cancel_lock);
 848 }
 849
 850 /*
 851  * scrub consumers
 852  */
 853
 854 static void
 855 count_block(zfs_all_blkstats_t *zab, const blkptr_t *bp)
 856 {
 857         int i;
 858
 859         /*
 860          * If we resume after a reboot, zab will be NULL; don't record
 861          * incomplete stats in that case.
 862          */
 863         if (zab == NULL)
 864                 return;
 865
 866         for (i = 0; i < 4; i++) {
 867                 int l = (i < 2) ? BP_GET_LEVEL(bp) : DN_MAX_LEVELS;
 868                 int t = (i & 1) ? BP_GET_TYPE(bp) : DMU_OT_TOTAL;
 869                 zfs_blkstat_t *zb = &zab->zab_type[l][t];
 870                 int equal;
 871
 872                 zb->zb_count++;
 873                 zb->zb_asize += BP_GET_ASIZE(bp);
 874                 zb->zb_lsize += BP_GET_LSIZE(bp);
 875                 zb->zb_psize += BP_GET_PSIZE(bp);
 876                 zb->zb_gangs += BP_COUNT_GANG(bp);
 877
 878                 switch (BP_GET_NDVAS(bp)) {
 879                 case 2:
 880                         if (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 881                             DVA_GET_VDEV(&bp->blk_dva[1]))
 882                                 zb->zb_ditto_2_of_2_samevdev++;
 883                         break;
 884                 case 3:
 885                         equal = (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 886                             DVA_GET_VDEV(&bp->blk_dva[1])) +
 887                             (DVA_GET_VDEV(&bp->blk_dva[0]) ==
 888                             DVA_GET_VDEV(&bp->blk_dva[2])) +
 889                             (DVA_GET_VDEV(&bp->blk_dva[1]) ==
 890                             DVA_GET_VDEV(&bp->blk_dva[2]));
 891                         if (equal == 1)
 892                                 zb->zb_ditto_2_of_3_samevdev++;
 893                         else if (equal == 3)
 894                                 zb->zb_ditto_3_of_3_samevdev++;
 895                         break;
 896                 }
 897         }
 898 }
 899
 900 static void
 901 dsl_pool_scrub_clean_done(zio_t *zio)
 902 {
 903         spa_t *spa = zio->io_spa;
 904
 905         zio_data_buf_free(zio->io_data, zio->io_size);
 906
 907         mutex_enter(&spa->spa_scrub_lock);
 908         spa->spa_scrub_inflight--;
 909         cv_broadcast(&spa->spa_scrub_io_cv);
 910
 911         if (zio->io_error && (zio->io_error != ECKSUM ||
 912             !(zio->io_flags & ZIO_FLAG_SPECULATIVE)))
 913                 spa->spa_scrub_errors++;
 914         mutex_exit(&spa->spa_scrub_lock);
 915 }
 916
 917 static int
 918 dsl_pool_scrub_clean_cb(dsl_pool_t *dp,
 919     const blkptr_t *bp, const zbookmark_t *zb)
 920 {
 921         size_t size = BP_GET_PSIZE(bp);
 922         spa_t *spa = dp->dp_spa;
 923         boolean_t needs_io;
 924         int zio_flags = ZIO_FLAG_SCRUB_THREAD | ZIO_FLAG_RAW | ZIO_FLAG_CANFAIL;
 925         int zio_priority;
 926
 927         ASSERT(bp->blk_birth > dp->dp_scrub_min_txg);
 928
 929         if (bp->blk_birth >= dp->dp_scrub_max_txg)
 930                 return (0);
 931
 932         count_block(dp->dp_blkstats, bp);
 933
 934         if (dp->dp_scrub_isresilver == 0) {
 935                 /* It's a scrub */
 936                 zio_flags |= ZIO_FLAG_SCRUB;
 937                 zio_priority = ZIO_PRIORITY_SCRUB;
 938                 needs_io = B_TRUE;
 939         } else {
 940                 /* It's a resilver */
 941                 zio_flags |= ZIO_FLAG_RESILVER;
 942                 zio_priority = ZIO_PRIORITY_RESILVER;
 943                 needs_io = B_FALSE;
 944         }
 945
 946         /* If it's an intent log block, failure is expected. */
 947         if (zb->zb_level == -1 && BP_GET_TYPE(bp) != DMU_OT_OBJSET)
 948                 zio_flags |= ZIO_FLAG_SPECULATIVE;
 949
 950         for (int d = 0; d < BP_GET_NDVAS(bp); d++) {
 951                 vdev_t *vd = vdev_lookup_top(spa,
 952                     DVA_GET_VDEV(&bp->blk_dva[d]));
 953
 954                 /*
 955                  * Keep track of how much data we've examined so that
 956                  * zpool(1M) status can make useful progress reports.
 957                  */
 958                 mutex_enter(&vd->vdev_stat_lock);
 959                 vd->vdev_stat.vs_scrub_examined +=
 960                     DVA_GET_ASIZE(&bp->blk_dva[d]);
 961                 mutex_exit(&vd->vdev_stat_lock);
 962
 963                 /* if it's a resilver, this may not be in the target range */
 964                 if (!needs_io) {
 965                         if (DVA_GET_GANG(&bp->blk_dva[d])) {
 966                                 /*
 967                                  * Gang members may be spread across multiple
 968                                  * vdevs, so the best estimate we have is the
 969                                  * scrub range, which has already been checked.
 970                                  * XXX -- it would be better to change our
 971                                  * allocation policy to ensure that all
 972                                  * gang members reside on the same vdev.
 973                                  */
 974                                 needs_io = B_TRUE;
 975                         } else {
 976                                 needs_io = vdev_dtl_contains(vd, DTL_PARTIAL,
 977                                     bp->blk_birth, 1);
 978                         }
 979                 }
 980         }
 981
 982         if (needs_io && !zfs_no_scrub_io) {
 983                 void *data = zio_data_buf_alloc(size);
 984
 985                 mutex_enter(&spa->spa_scrub_lock);
 986                 while (spa->spa_scrub_inflight >= spa->spa_scrub_maxinflight)
 987                         cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
 988                 spa->spa_scrub_inflight++;
 989                 mutex_exit(&spa->spa_scrub_lock);
 990
 991                 zio_nowait(zio_read(NULL, spa, bp, data, size,
 992                     dsl_pool_scrub_clean_done, NULL, zio_priority,
 993                     zio_flags, zb));
 994         }
 995
 996         /* do not relocate this block */
 997         return (0);
 998 }
 999
1000 int
1001 dsl_pool_scrub_clean(dsl_pool_t *dp)
1002 {
1003         /*
1004          * Purge all vdev caches.  We do this here rather than in sync
1005          * context because this requires a writer lock on the spa_config
1006          * lock, which we can't do from sync context.  The
1007          * spa_scrub_reopen flag indicates that vdev_open() should not
1008          * attempt to start another scrub.
1009          */
1010         spa_config_enter(dp->dp_spa, SCL_ALL, FTAG, RW_WRITER);
1011         dp->dp_spa->spa_scrub_reopen = B_TRUE;
1012         vdev_reopen(dp->dp_spa->spa_root_vdev);
1013         dp->dp_spa->spa_scrub_reopen = B_FALSE;
1014         spa_config_exit(dp->dp_spa, SCL_ALL, FTAG);
1015
1016         return (dsl_pool_scrub_setup(dp, SCRUB_FUNC_CLEAN));
1017 }